Learn
Explainers, standards, and guidance on AI agent risk and shadow AI, so the claims on this site have a second source.
AI Inception cites the OWASP Top 10 for Agentic Applications, which lists rogue AI agents as ASI10.
Watch
Eleven explainers from IBM Technology, OWASP, Anthropic, Google Cloud, and RSAC, two at a time, with the player loading only when you press play.
1 to 2 of 11
Read
The standards, frameworks, and guidance this site leans on, each one linked to its source.
Standards and frameworks
- OWASP GenAI Security ProjectOWASP Top 10 for Agentic Applications
The ten risks specific to AI agents, with rogue AI agents listed as ASI10, which this site cites.
- OWASP GenAI Security ProjectOWASP Top 10 for LLM Applications 2025
The ten risks for applications built on large language models, from prompt injection to excessive agency.
- OWASP GenAI Security ProjectAgentic AI – Threats and Mitigations
A threat model of emerging agentic threats and the mitigations the OWASP Agentic Security Initiative recommends.
- NISTAI Risk Management Framework
The voluntary framework for managing AI risk through its govern, map, measure, and manage functions.
- MITREMITRE ATLAS
A knowledge base of adversary tactics and techniques against AI systems, modeled on ATT&CK.
- GoogleGoogle's Secure AI Framework
A practitioner guide to AI security with controls and a self assessment for your own deployment.
Guidance from government agencies
- UK National Cyber Security CentreGuidelines for secure AI system development
Secure design, development, deployment, and operation guidance agreed by cyber security agencies across many countries.
- CISADHS CISA and UK NCSC Release Joint Guidelines for Secure AI System Development
The announcement of the joint guidelines and the secure by design principles behind them.
From the companies building and securing AI agents
- AnthropicBuilding effective AI agents
Anthropic on the simple, composable patterns behind the AI agents that work, and when not to build one.
- Microsoft SecurityNew whitepaper outlines the taxonomy of failure modes in AI agents
The Microsoft AI Red Team on the ways AI agents fail across safety and security, novel and familiar.
- Cloud Security AllianceAgentic AI Threat Modeling Framework: MAESTRO
A threat modeling method built for systems of many AI agents, layer by layer, from the Cloud Security Alliance.
- IBMWhat Is Shadow AI?
The IBM explainer on AI tools adopted without approval and why they matter for security and governance.
- IBMWhat Are AI Agents?
A plain definition of AI agents, how they differ from assistants, and the kinds in use today.
Every link on this page was checked on 22 October 2026.
The words on this site
The words used across this site, in the sense we mean them.
- AI agent
- Software that pursues a task on its own by calling models, tools, and data, instead of answering one prompt.
- Scope
- The role, task, data, and tools an AI agent was approved for, written down so a step outside it is visible.
- Drift
- The first step an AI agent takes away from its scope, often small and easy to miss on its own.
- Restore
- Returning an AI agent and what it touched to the last state a person approved.
- Revoke
- Taking away an access or a tool an AI agent should not hold again, and telling a person.
- Collusion
- Two or more AI agents coordinating outside their roles while each one looks legitimate on its own.
- Rogue AI agent
- The OWASP name, ASI10, for an AI agent that deviates from its intended purpose, including through collusion.
- Shadow AI
- AI tools and AI agents people adopt for work without approval, so nobody is watching what they reach.
- AI inventory
- One list of every AI tool and AI agent in use, where each runs, and who approved it.
- Human on the loop
- A person who sees what an AI agent does as it happens and decides the steps that must wait.
- Prompt injection
- Instructions hidden in content an AI agent reads, written to make it act against its task.
Questions that come up first
Short answers to the questions that come up first on a demo call.
What is the difference between an AI agent and a chatbot?
A chatbot answers the prompt in front of it and then waits for the next one. An AI agent pursues a task over many steps, calling tools and data along the way.
Why does drift matter if the AI agent still finishes the job?
The job can finish while a step along the way touched data or a system nobody approved. The record of that step is what lets a person judge it.
Is shadow AI the same as shadow IT?
Shadow IT is software people adopt without approval, and shadow AI is the same pattern with AI tools and AI agents. The difference is that an AI agent can act, so the risk is not only data leaving.
What does human on the loop mean in practice?
A person does not approve every step, which would slow the AI agent to a crawl. Instead the person sees what happens, sets which actions wait, and decides those.
How does AI Inception relate to the OWASP Top 10 for Agentic Applications?
AI Inception is built around the risks that list names, with rogue AI agents and collusion front and center. The list is the outside reference for the claims on this site.
Which AI agents can be watched?
Coding, sales, cloud, and local AI agents, whatever framework or model provider runs them, in the cloud or on your own hardware.
See AI Inception on your own AI agents
Tell us where your AI agents run and we will schedule a demo on a live console with your team.
Request a demo